The long tail of the category tree
Category size against rank, from one article to 1.1 million
Wikipedia's categories were never a shelving scheme: an article can sit in a dozen at once, and nothing stops a category from ballooning. Sort the 1 834 777 categories of release 2606 by how many articles each directly holds and the distribution collapses into a heavy head and a very long tail — one category carries more than a million memberships while hundreds of thousands carry exactly one article.
Source: K3-catpyramid receipt (B2 category long tail), via the dataset page.
The heaviest head
The receipt names the heaviest bucket: the living-people category, whose 1 135 723 direct articles are 2.7% of all memberships on their own. Concentration compounds quickly from there. The ten largest categories hold 4.0% of all memberships, the top hundred 7.5%, the top thousand 16.1% — and the top ten thousand, half a percent of the tree, hold 31.4%.
The tail is the mirror image: 295 547 categories (16.1% of categories) hold exactly one article and 888 113 (48.4%) hold between two and nine, so 64.5% of the tree sits below ten articles. The median category is nearly empty; the mass lives in the head.
What the fit does and does not say
A near-straight line in log-log space is the Zipf signature: size proportional to a power of rank. The receipt's OLS fit over the top 10 000 categories gives slope −0.73 with r = −0.998 — but the fitted slope steepens as the window widens (−0.55 over the top 100, −0.69 over the top 1 000), so this is a Zipf-like regularity of the head, not a law of the whole tree. The figure draws the amber line only across its fitted window and never extrapolates it into the tail. Nothing hides at the floor, either: the receipt records 0 categories with zero direct articles, so the one-article band is the true bottom of the tree.
How to count a membership
A membership is one direct incidence: an article sitting in five categories contributes five memberships, so the 41.4M total is incidence, not distinct-article coverage — categories overlap, and the receipt documents the double-counting. Direct also excludes the tree's parent edges: a subcategory's contents are not folded into its ancestors; each category is banded by its own article count. Single-category bands are drawn as points at their exact size; multi-category bands join corner to corner because the payload bins the counts, so each band's endpoints, counts, and totals are exact while its interior path is schematic.
The full receipted payload — the eight bands, the top-N shares, all three Zipf fits, and the zero-article count — is the figure's data payload, described on the dataset page.